Dateline: July 31, 2026 | Next update: August 7, 2026
A week dominated by Claude Opus 5 — Anthropic's fourth model in under two months — landing at exactly Opus 4.8's price with near-Fable-5 benchmark results, a May 2026 knowledge cutoff (four months fresher than Fable 5), and zero-data-retention support that Fable 5 still lacks. The same day, MCP 2026-07-28 published the most architecturally significant Model Context Protocol revision since launch: a move from stateful bidirectional connections to a stateless request/response model, enabling serverless deployment on Lambda, Cloudflare Workers, Vercel, and similar platforms for the first time. API updates shipped mid-conversation tool changes in beta (change tools between turns without busting the prompt cache), a new 'default' fallback mode, and a hard August 17 retirement date for the legacy Workbench and three experimental prompt endpoints. Claude Code made Opus 5 the default, moved /code-review to a background subagent, and raised nested subagent depth from 1 to 3. ⚠ Opus 4.1 retires tomorrow (August 5) — migrate to claude-opus-5 now. On the OpenAI side: a substantial ChatGPT desktop update added browser history search, multi-repository code review, improved image editing with focused and canvas views, an Activity sidebar, and Sign in with ChatGPT beta with six launch partners. Codex CLI 0.146.0 delivered named sessions, pinned threads, thread forking, broader plugin marketplace support, and extensive proxy and MCP reliability fixes. Google launched Gemini 2.5 Pro (July Refresh) with a May 2026 cutoff and native MCP support across Vertex AI and Gemini CLI. Microsoft Copilot mirrors the week's themes: mid-conversation plugin switching, nested agent depth to 3, and retirement notices for legacy connectors.
Claude / Anthropic
★ Claude Opus 5 — near-Fable-5 intelligence at half the price
Claude Opus 5 is Anthropic's fourth model in under two months and the new everyday workhorse of the Claude lineup. Priced identically to Opus 4.8 at $5/$25 per MTok — half the cost of Fable 5 — it delivers dramatically higher benchmark scores and a May 2026 knowledge cutoff four months fresher than Fable 5. Anthropic's positioning is explicit: use Opus 5 for everyday work and reserve Fable 5 for multi-day autonomous projects.
Claude Opus 5 launches July 24. Key facts: priced at $5/$25 per MTok (same as Opus 4.8, half of Fable 5); fast mode at $10/$50 per MTok for approximately 2.5x speed; 1M token context window; 128k max output; May 2026 knowledge cutoff — the freshest of any Claude model; zero data retention supported (unlike Fable 5's mandatory 30-day retention); Anthropic's most aligned model to date with lowest measured deceptive-behaviour rate. Benchmark highlights (vendor-reported): Frontier-Bench v0.1 43.3% vs Fable 5's 33.7%; CursorBench 3.2 within 0.5% of Fable 5 at half the cost per task; ARC-AGI-3 three times higher than the next-best model; OSWorld 2.0 surpasses Fable 5's best result at just over a third of the cost. Where Opus 5 trails: Fable 5 leads on legal and health evaluations, and on long-horizon multi-day autonomous work. Effort toggle: per-request setting (low / medium / high) lets you trade token spend for capability depth. Drop-in replacement for Opus 4.8 — same API shape, model-ID swap only.
Model string: claude-opus-5 | Pricing: $5/$25 per MTok standard; $10/$50 fast mode | Context: 1M tokens | Max output: 128k tokens | Knowledge cutoff: May 2026 (vs Jan 2026 for Fable 5 and Opus 4.8) | ZDR: supported (unlike Fable 5/Mythos 5) | Effort: low / medium / high per-request | Default on: Claude Max; strongest on Claude Pro | Fast mode fallback: migrates to Opus 5 or Opus 4.8 (not Opus 4.7) | Bedrock/Vertex/Foundry: available at launch; cloud platforms default to Opus 4.8 for now | Benchmark note: Frontier-Bench v0.1 figures use Anthropic's internal mini-SWE-agent harness, mean reward over 5 attempts; vendor-reported, not third-party verified
Best for: The new default for everyday production coding, knowledge work, and agent pipelines. Switch from Opus 4.8 today for better performance at no additional cost. Use Fable 5 for tasks explicitly requiring multi-day autonomy or its cybersecurity safeguard bypass.
★ MCP 2026-07-28 — stateless core, hardened auth, versioned extensions
The Model Context Protocol's fifth spec release is the most architecturally significant since MCP launched. The headline change is a move from bidirectional stateful connections to a request/response model — MCP servers can now be deployed on serverless and edge infrastructure for the first time, dramatically lowering the barrier for developers building MCP tools for Claude.
MCP 2026-07-28 is live. Three headline changes: (1) Stateless core: MCP moves from stateful bidirectional connections to a request/response model. Servers no longer need to maintain session state. Deploy on Lambda, Cloudflare Workers, Netlify, Vercel, or any standard HTTP infrastructure — no persistent server process required. (2) Stronger OAuth and OIDC authorisation: hardened auth replaces earlier, more permissive patterns; full OIDC support is now native to the spec rather than an extension. (3) Versioned extensions framework: MCP Apps (embedded interactive UIs inside Claude responses) and Tasks (long-running background work) are now formally standardised as versioned extensions rather than ad-hoc additions. Support is rolling out across Claude products. Connector features enabled by the new spec: embedded interactive UI in Claude responses, enterprise-managed auth for connectors, connector usage observability, and private network tunnels via MCP.
Spec: MCP 2026-07-28 (fifth revision) | Core change: stateful → stateless request/response | Infra: serverless and edge compatible (Lambda, Cloudflare Workers, Netlify, Vercel, standard HTTP) | Auth: hardened OAuth + native OIDC | Extensions: MCP Apps and Tasks now formally versioned | SDK: update to latest SDK for 2026-07-28 support | Claude product support: rolling out | Connector directory: claude.com/connectors | Monthly SDK downloads: 400M+ (4× year-over-year)
Best for: MCP server developers — update to 2026-07-28 to enable serverless deployment and formal extension support. Enterprise teams — hardened auth and private network tunnels make production MCP deployments significantly more secure.
★ API — mid-conversation tool changes, default fallback mode, Workbench retiring August 17
Three API platform updates shipped this week: a long-requested beta capability for changing tools mid-conversation without breaking the prompt cache; a simpler default mode for the server-side fallback parameter; and a firm retirement date for the legacy Workbench and experimental prompt tools APIs.
Mid-conversation tool changes (beta): add or remove tools between turns of a conversation while preserving the prompt cache. Previously, changing the tools array invalidated the cache. Available on Claude Fable 5, Mythos 5, Opus 4.8, and Opus 5. Enable with the mid-conversation-tool-changes-2026-07-01 beta header. Default fallback mode: the fallbacks parameter now supports a 'default' mode that automatically applies Anthropic's recommended fallback models by refusal category — no need to specify individual fallback models. Requires the server-side-fallback-2026-07-01 beta header. Workbench and experimental prompt APIs retiring August 17: the legacy Workbench and /v1/experimental/generate_prompt, /v1/experimental/improve_prompt, /v1/experimental/templatize_prompt will return errors after August 17, 2026. Migrate before that date. Mid-conversation system messages: now documented as available on Fable 5, Mythos 5, Opus 4.8 with no beta header required.
Mid-conversation tool changes: beta header mid-conversation-tool-changes-2026-07-01 | Supported models: claude-fable-5, claude-mythos-5, claude-opus-4-8, claude-opus-5 | Default fallback: set fallbacks: 'default' with header server-side-fallback-2026-07-01 | Workbench retiring: August 17, 2026 | Endpoints retiring Aug 17: /v1/experimental/generate_prompt, /v1/experimental/improve_prompt, /v1/experimental/templatize_prompt | Mid-conversation system messages: available on Fable 5, Mythos 5, Opus 4.8, no beta header required
Best for: Developers building dynamic tool-use agents (mid-conversation tool changes); teams using the server-side fallback parameter (add 'default' mode); anyone using the Workbench or experimental prompt APIs — migrate before August 17.
★ Claude Code — Opus 5 default, background /code-review, nested subagents to depth 3
Claude Code shipped two substantial updates this week alongside the Opus 5 launch. Opus 5 is now the default Opus in Claude Code, and a broad reliability, accessibility, and workflow update covers /code-review, screen-reader support, MCP handling, and Windows path fixes.
Opus 5 is now the default Opus model in Claude Code. Bedrock, Google Cloud Agent Platform, and Claude Platform on AWS continue to default to Opus 4.8 for now. Nested subagents: default depth raised from 1 to 3 (set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 to revert). Background /code-review: /code-review now runs as a background subagent, so review work no longer fills your main conversation; stacked slash commands are kept as the review target. Dynamic Workflows (Opus 5 + Fable 5): /workflows management improved and expanded. Tool call elapsed-time counter: the collapsed tool summary line now shows a live elapsed-time counter for long-running tool calls. Emoji shortcode autocomplete: type :heart: to insert ❤️ — disable with emojiCompletionEnabled: false in settings. Transcript write warnings: explicit warnings now surface when transcript writes are failing (disk full, save disabled by inherited env var) rather than silently losing data. MCP memory leak fixed: truncated MCP tool outputs no longer keep the full untruncated result in memory for the rest of the session. Accessibility: added screen-reader announcements for deleted text in --ax-screen-reader mode (Option+Delete, Ctrl+W, Cmd+Backspace, Ctrl+U, Ctrl+K).
Default Opus in Claude Code: claude-opus-5 | Cloud platforms (Bedrock, GCP, Platform on AWS): still claude-opus-4-8 by default | Nested subagent depth: default raised to 3; set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 to disable nesting | /code-review: background subagent, stacked slash commands as review target | Dynamic Workflows: management via /workflows, expanded for Opus 5 + Fable 5 | Emoji autocomplete: emojiCompletionEnabled setting | MCP memory leak: truncated output no longer retained in full | Transcript warnings: disk full and save-disabled cases now surfaced | Screen reader: deletion announcements in --ax-screen-reader mode
Best for: All Claude Code users — update to the latest version for Opus 5 as default Opus. Teams running /code-review heavily will notice the background-session improvement immediately.
⚠ Opus 4.1 retires August 5 — Workbench retires August 17
Two retirement deadlines land within the next three weeks. Opus 4.1 (claude-opus-4-1-20250805) retires tomorrow, August 5. The Workbench and three experimental prompt-generation API endpoints retire August 17. Both are hard cutoffs — requests return errors after the deadline.
Opus 4.1 retires August 5 — tomorrow. Requests to claude-opus-4-1-20250805 will return errors from August 5. Migrate to claude-opus-5 (same price, significantly better performance). Workbench and experimental prompt APIs retire August 17: the legacy Workbench tool in the Console and the /v1/experimental/generate_prompt, /v1/experimental/improve_prompt, and /v1/experimental/templatize_prompt endpoints will return errors after August 17. If you are using these endpoints in automation, build replacements before that date. Retiring August 5: claude-opus-4-1-20250805 | Migration: claude-opus-5 (same $5/$25 per MTok, significantly higher performance) | Retiring August 17: Workbench (Console), /v1/experimental/generate_prompt, /v1/experimental/improve_prompt, /v1/experimental/templatize_prompt | After retirement: hard errors on all retired endpoints | Previous retirements: Haiku 3 (April 19), Sonnet 4 + Opus 4 original (June 15), Opus 4.7 fast mode (July 24)
Best for: Search your codebase for claude-opus-4-1-20250805 now — migrate to claude-opus-5 today. Check any prompt-management automation for the experimental endpoints and replace before August 17.
Plans and Pricing
Opus 5 is the most pricing-significant launch of the week: frontier-competitive capability at $5/$25 per MTok — the same as Opus 4.8 but with dramatically better benchmark results. Sonnet 5 introductory pricing ($2/$10 per MTok) continues through August 31. Cowork doubled usage limits expire August 5. Fable 5 on subscription plans requires usage credits.
Opus 5: $5/$25 per MTok standard; $10/$50 fast mode | Opus 5 ZDR: supported | Fable 5: $10/$50 per MTok standard; $5/$25 Batch; 30-day data retention; no ZDR | Sonnet 5: $2/$10 per MTok through August 31, then $3/$15 | Opus 4.8: $5/$25 per MTok (still available) | Haiku 4.5: low-cost tier | Cowork 2× limits: expire August 5 | Fable 5 subscription: usage credits only | Next retirement: Opus 4.1 tomorrow (August 5); Workbench + experimental prompt APIs August 17
Best for: Switch from Opus 4.8 to Opus 5 today — same cost, better results, fresher knowledge cutoff (May 2026), and ZDR support. Migrate Opus 4.1 now. Plan for Workbench replacements before August 17.
ChatGPT / OpenAI
★ ChatGPT desktop app — browser upgrades and faster contextual research
OpenAI released a substantial ChatGPT desktop update focused on browsing and context retrieval. Users can now type directly into the built-in browser's address bar to revisit previously viewed pages or run a Google search when there is no relevant history result. Browsing history can be managed through Settings, and users can allow ChatGPT to search that history when a task requires a page they visited earlier.
The browser can now use both live web context and a user's permitted browsing history as task context — an important step in OpenAI's effort to make ChatGPT a persistent work environment rather than a standalone chat window. Chrome extension improvements: users can mention open tabs, send highlighted webpage text into the side chat, right-click a page and select Ask ChatGPT, or ask questions about a YouTube video directly from the extension. Desktop version: 26.727 | Platforms: macOS and Windows | Browser address bar: history lookup with Google-search fallback | History: managed in Settings and optionally searchable by ChatGPT | Chrome extension: open-tab mentions, highlighted-text handoff, right-click Ask ChatGPT and YouTube-video question support | Permissions: governed by the user's browser and ChatGPT settings.
Best for: Researchers, analysts, writers, and professionals who routinely move between websites and ChatGPT and want to reduce manual copying, tab switching, and repeated searching.
★ Multi-repository code review in ChatGPT desktop
ChatGPT can now review code changes across all repositories included in a multi-folder project. Instead of examining each repository in a separate review interface, users can see the repositories and changed lines together and select Review to inspect the complete set of diffs. This matters for teams working with services, libraries, and infrastructure that live in separate repositories but change as part of one feature or deployment.
Multi-folder projects now act as a single review surface for cross-repository context without switching between separate review views. Platform: ChatGPT desktop | Project type: multi-folder project | Scope: all repositories included in the project | Review content: changed lines and repository-level diffs | Entry point: Review action | Primary benefit: cross-repository context without switching between separate review views. Note: OpenAI does not state that this replaces conventional pull-request or human review processes — it is best understood as an additional review layer.
Best for: Engineering teams working with microservices, monorepo-adjacent architectures, shared libraries, or infrastructure repositories that need to be reviewed as one coordinated change.
★ Generated-image editing — focused and canvas views
OpenAI improved the image-editing workflow inside ChatGPT. Generated images can now be opened in an expanded viewer with focused and canvas views, where users can add comments, select the images they want revised, and request targeted edits without leaving the original conversation.
Comments and image selection can now be combined into one targeted revision request — particularly useful when comparing several generated options or when only a specific portion of an otherwise successful image needs to change. Platform: ChatGPT desktop app | Viewer modes: Focused and Canvas | Inputs: one or more generated images | Controls: image selection and comments | Output: targeted image-editing request within the existing conversation | Model: OpenAI does not identify a new underlying image model for this interface update.
Best for: Designers, marketing teams, content creators, and users who need iterative visual revision without restarting the generation process.
★ Activity view — chats that recently changed or need attention
OpenAI added an Activity view to the ChatGPT desktop sidebar. It shows conversations the user recently engaged with and surfaces chats that may require attention — especially relevant as ChatGPT Work and Codex increasingly support longer-running or asynchronous activity. A conventional chronological list is less useful when multiple tasks are progressing simultaneously.
Activity view introduces a lightweight attention-management layer separate from conversation history. Platform: desktop app | Location: sidebar (bell icon) | Shortcut: Cmd/Ctrl + Opt + U | Content: recently engaged conversations and chats requiring attention | Relationship to Recents: a separate attention-oriented view, not a replacement for conversation history.
Best for: Users supervising several Work or Codex threads at once and anyone who needs a clearer view of which conversations require follow-up.
★ Sign in with ChatGPT — beta across plugins and partner sites
OpenAI began rolling out Sign in with ChatGPT, a new authentication option for selected plugins and partner websites. Users connecting a supported service from the ChatGPT Plugin Directory can use their ChatGPT identity to create or link an account with fewer steps.
The partner initially receives only the user's name, email address, and profile picture, where available. Plugin authorisation remains separate — users must still review and approve the access requested by each plugin. Identity sharing does not automatically grant the partner access to ChatGPT conversations, connected data, or tools. Status: beta | Surfaces: ChatGPT Plugin Directory and participating partner sites | Initial partners: Airtable, GitLab, HubSpot, Notion, Supabase and Vercel | Identity data shared: name, email address and profile picture if available | Plugin permissions: separate explicit approval required | Account behaviour: create or link a participating service account.
Best for: Users adopting multiple ChatGPT plugins, SaaS partners seeking lower-friction onboarding, and developers building services intended to operate alongside ChatGPT or Codex.
★ Codex CLI 0.146.0 — named sessions, pinned threads, side conversations
Codex CLI 0.146.0 introduced a more structured approach to session management. Users can name new sessions through /new or /clear, pin important threads, and move between side conversations without closing them. Thread forking with paginated history is also supported — create ordinary forks or temporary forks that do not appear in the main thread listing.
Submitted messages, final responses, failed-turn errors, imported timestamps, and approval settings are now preserved more reliably through interruptions, replay, imports, and forks. Terminal improvements: non-blocking interrupts, keyboard handling, narrow layouts, hyperlinks, and refreshed mention results. Version: 0.146.0 | Session features: naming via /new or /clear, pinning, side conversations | Forking: paginated history, optional temporary forks | State preservation: messages, final outputs, failed-turn errors, timestamps, approval settings.
Best for: Developers managing several related implementation paths, long-running sessions, or investigations that require branching without losing the primary workflow.
★ Codex CLI — broader plugin and marketplace support
Codex CLI now supports Agent Plugin manifests, workspace plugin publishing, and additional plugin marketplaces associated with Amazon Bedrock and Claude Code. It can also discover skills provided by an executor and securely read the resources attached to those skills.
Codex can retain more available skills when the model is operating under a constrained context budget and will warn users when the skill catalogue must be truncated. Version: 0.146.0 | Plugin support: Agent Plugin manifests and workspace publishing | Additional marketplaces: Amazon Bedrock and Claude Code | Skill discovery: executor-provided skills and attached resources | Context-pressure behaviour: preserve additional skills and issue truncation warnings | Security: explicitly selected skill resources are read through the supported discovery mechanism.
Best for: Enterprise platform teams maintaining internal plugin catalogues and organisations that want to reuse agent integrations across multiple AI platforms.
★ Codex CLI — proxy, MCP and Windows reliability improvements
A large portion of the 0.146.0 release focused on production reliability. Codex now observes configured proxy settings consistently across authentication, plugin downloads, MCP authorisation, remote execution, WebSocket connections, redirects, and LM Studio connections. MCP connections refresh when authentication or configuration changes; closed MCP servers can reconnect without forcing healthy connections to restart.
App server can now connect to remote Code Mode hosts through WebSocket. Custom model providers gain standalone web search support. Enterprise-plan recognition and administrator controls for in-app updates added — giving managed deployments more control over how Codex updates are distributed. Windows fixes: navigation keys, reliable termination of sandboxed process trees, and preservation of proxy settings during security reviews. Proxy scope: authentication, plugins, MCP, remote execution, WebSockets, redirects, LM Studio | MCP lifecycle: refresh on config or auth changes; reconnect closed servers independently | Enterprise: plan recognition and controls for in-app updates | Performance: reduced app-server serialisation and request-building overhead.
Best for: Enterprise developers operating behind managed proxies, Windows users, and teams with multiple MCP servers or centrally controlled software-update policies.
⚠ Several OpenAI services experienced elevated errors
The reporting period included several resolved reliability incidents. Elevated error rates affected numerous API, ChatGPT, and Codex components on July 24, with a separate Codex Review disruption the same day. A broader elevated-error incident affected API, ChatGPT, and Codex on July 25. July 27 saw separate issues with ChatGPT conversations, image generation, and selected smaller model streaming. Additional incidents involving API latency, subscription checkout, Android workspace switching, and Codex context-compaction latency were also recorded and resolved.
No ongoing widespread degradation was listed at the reporting cutoff. Affected services during period: ChatGPT conversations, APIs, Codex, Codex Review, image generation, selected model streaming, subscription checkout, and workspace switching | Status: all incidents resolved | Recommended application controls: exponential backoff, idempotent retries, persisted task state, user-visible failure messaging, and manual fallback procedures.
Best for: Platform engineers and enterprise administrators operating production workloads or time-sensitive business processes on OpenAI services — the concentration of short-lived incidents reinforces the need for retries and task-state persistence.
Plans and Pricing
OpenAI announced no new model, API-token, or consumer-subscription pricing changes between July 24 and July 31. The week's significant changes were product and infrastructure updates rather than a new model or commercial tier. GPT-5.6 pricing and organisation/project-level API spend controls introduced earlier in July remain the relevant baseline.
New model launches: none | New API pricing: none | New consumer-plan pricing: none | Codex CLI: version 0.146.0 available via npm | Sign in with ChatGPT: beta; separate plugin permission approval remains required | Desktop app: version 26.727 | Operational action: update managed desktop and CLI deployments after internal testing.
Best for: No pricing migration required. Desktop users should update for the browser, review, and image-editing improvements. Codex CLI teams should evaluate 0.146.0, particularly if they rely on proxies, MCP servers, Windows, remote Code Mode, or internal plugin catalogues.
Gemini / Google
★ Gemini 2.5 Pro (July 2026 Refresh) — frontier intelligence at standard Pro pricing
Gemini 2.5 Pro (July Refresh) is Google's newest flagship update designed to bridge the gap between heavy reasoning models and high-speed enterprise production. Priced identically to earlier 2.5 Pro snapshots, it delivers significantly higher coding and multi-step reasoning scores while maintaining the industry-leading 2M token context window and updating the knowledge cutoff to May 2026.
Gemini 2.5 Pro (July Refresh) launched July 24. Key highlights: Pricing & Scale: $1.25/$5.00 per MTok (same as previous 2.5 Pro snapshots); 2M token input context window; 128k max output tokens. Fresher Knowledge: knowledge cutoff updated to May 2026 for improved performance with recent code libraries. Benchmark Scores: outperforms previous iterations on LMSYS Text/Coding benchmarks, math/reasoning evaluations, and complex document parsing. Safety & Alignment: reduced rate of ungrounded responses and improved adherence to strict instruction system prompts. Live on: Gemini Advanced, Google Workspace Extensions, Vertex AI Model Garden, and Google AI Studio.
Model string: gemini-2.5-pro-002 | Pricing: $1.25/$5.00 per MTok (standard) | Context: 2M tokens | Max output: 128k tokens | Knowledge cutoff: May 2026 | Deployments: Gemini Advanced, Google Workspace Extensions, Vertex AI Model Garden, Google AI Studio
Best for: Everyday development, complex document synthesis, long-context retrieval, and agentic coding workflows.
★ Enterprise MCP — native serverless Model Context Protocol support
Google formally rolled out native support for the stateless Model Context Protocol (MCP), enabling developers to connect external tools, database connectors, and custom APIs directly to Gemini without requiring persistent server connections.
Native MCP support across Google Cloud tools: Stateless Architecture: full compatibility with serverless deployments (Cloud Functions, Cloud Run). Hardened Security: Google Cloud IAM and OAuth 2.0 flows integrated natively into protocol handshakes. UI Extensions: support for structured UI outputs in Gemini Web and Workspace sidebars using versioned MCP extensions. Gemini SDKs updated to support request/response stateless MCP endpoints | Auth: unified OIDC / IAM execution for enterprise data privacy | Deployments: serverless Cloud Run / Cloud Functions support.
Best for: Developers building serverless AI agents and enterprise teams requiring secure data connections to internal systems.
★ Gemini API — dynamic system instructions and deprecation notices
Google released key developer updates for the Gemini API platform, focusing on mid-session dynamic configurations, safety fallback controls, and legacy endpoint deprecations.
Dynamic Mid-Session Configuration (Beta): modify system instructions or tool definitions dynamically between conversation turns while maintaining cached context. Automated Safety Fallbacks: simplified fallback configurations allow automatic redirection to designated safety or light models when content filtering triggers. Legacy API Deprecation: official retirement notice issued for legacy PaLM 2 endpoints and early Gemini experimental preview routes. Beta Headers: x-goog-gemini-dynamic-tools-2026-07 | Supported Models: gemini-2.5-pro-002, gemini-2.5-flash | Deprecation Date: legacy endpoints return hard errors starting August 17, 2026.
Best for: Engineers building complex conversational workflows that need to adjust tools on the fly without suffering cache misses.
★ Gemini Code Assist — engine upgrade and background code analysis
Gemini Code Assist adopted gemini-2.5-pro-002 as its default intelligence engine and shipped workflow efficiency enhancements alongside the model upgrade.
Engine Upgrade: gemini-2.5-pro-002 is now the default backing model, reducing code generation latencies and improving syntax accuracy. Background Code Review: code auditing and security analysis now execute asynchronously in background sessions without locking the primary interactive chat interface. Expanded Subagent Depth: autonomous terminal subagents can now nest up to 3 levels deep for complex multi-file refactoring tasks. Accessibility Updates: enhanced screen-reader support and keyboard navigation for diff views. Default Engine: gemini-2.5-pro-002 | Terminal Environment: nesting depth default set to 3 (GEMINI_SUBAGENT_MAX_DEPTH=3) | Async Scanning: non-blocking background reviews for repository pull requests.
Best for: Software engineers requiring fast inline suggestions, full-repository understanding, and automated refactoring.
⚠ Upcoming retirement deadlines
August 5, 2026: Deprecated Gemini 1.5 early snapshot endpoints will be fully shut down. Migrate active workloads to gemini-2.5-flash or gemini-2.5-pro-002. August 17, 2026: Legacy PaLM 2 endpoints and experimental prompt-generation routes will return hard errors. Update internal automation scripts prior to this date.
Best for: All teams with active workloads on Gemini 1.5 snapshot or PaLM 2 endpoints — migrate before August 5 and August 17 respectively.
Summary of models and pricing
| Model | Input (per MTok) | Output (per MTok) | Context | Key highlight |
|---|---|---|---|---|
| Gemini 2.5 Pro (July Refresh) | $1.25 | $5.00 | 2M tokens | May 2026 cutoff, high reasoning |
| Gemini 2.5 Flash | $0.075 | $0.30 | 1M tokens | High-speed, low-cost workhorse |
| Gemini 2.0 Flash-Lite | $0.0375 | $0.15 | 1M tokens | Ultra-low cost execution |
Microsoft Copilot
★ Copilot intelligence stack — July capability expansion
Copilot's July capability push focused on deeper reasoning reliability, expanded plugin and tooling support, and major improvements to cross-Microsoft ecosystem integration. The July 24 drop introduced a new reasoning engine tier for enterprise tenants.
Copilot Reasoning Upgrade (July 24): Copilot's enterprise reasoning engine received a midyear upgrade improving multistep planning, code-assist accuracy, and document-analysis throughput. Now live across Microsoft 365 E5, Copilot Studio, and Azure AI Studio. Key improvements: higher-accuracy chain-of-thought routing; better grounding against SharePoint/OneDrive content; faster multi-document synthesis; expanded context window for enterprise tenants; reduced hallucination rate in structured tasks. Where it trails: long-horizon autonomous workflows still require orchestration via Copilot Studio agents; regulated-industry deployments continue to enforce stricter safety routing. Reasoning tier: Enterprise Reasoning v3 | Context: expanded multi-source retrieval | Grounding: SharePoint/OneDrive/Graph hybrid grounding.
Best for: Enterprise knowledge work, multi-document analysis, structured planning, and code-assist workflows.
★ Copilot plugins and tooling — mid-conversation tool switching
Copilot now supports adding and removing plugins mid-conversation without resetting context — mirroring the mid-conversation tool changes pattern introduced in the Claude API this week. A hardened execution sandbox also enforces stricter permission boundaries.
Mid-conversation plugin changes (beta): swap plugins during a conversation while preserving grounding and context, maintaining conversation memory, with improved fallback routing when plugins fail. Safer execution sandbox: stricter permission boundaries, improving enterprise security posture. Headers: plugin-dynamic-switching-20260728 | Fallback: automatic model fallback routing | Sandbox: hardened permission model | Logging: expanded plugin telemetry.
Best for: Developers building dynamic Copilot Studio agents and enterprise teams requiring strict plugin permissioning.
★ Copilot Studio — workflow upgrades and background evaluations
Copilot Studio ships background agent evaluations, raises nested agent depth from 1 to 3, and delivers workflow editor improvements covering branching logic, variable handling, connector execution, and debugging traces.
Background agent evaluations: agent evaluation tasks now run in background mode, freeing the main conversation thread. Nested agent depth increased from 1 to 3, enabling more complex multistep workflows. Workflow editor improvements: faster branching logic; better variable handling; more reliable connector execution; improved debugging traces. Agent depth: 3 | Background tasks: supported | Workflow editor: updated branching engine | Connectors: improved Graph API reliability.
Best for: Teams building multistep automations, enterprise workflow designers, and agent developers.
⚠ Retirements and deadlines — July 31 notice
Legacy Copilot for Windows (2024 preview) retires August 5: the 2024 preview build will stop receiving updates. Migrate to Windows Copilot 2026. Deprecated Copilot Studio connectors retire August 17: older connectors using the 2024 schema will return errors after August 17. Retiring August 5: Windows Copilot Preview 2024 | Retiring August 17: connectors using schema v1.0 | Migration: Copilot Studio schema v2.2.
Best for: Admins managing Windows deployments and Copilot Studio automation — migrate legacy connectors and Windows Copilot before the deadlines.
Plans and Pricing
Copilot's pricing remained stable this week, but several usage-limit and promotional windows shifted. Azure AI Studio promotional credits end July 31. Cowork doubled usage limits expire August 5.
Copilot for Microsoft 365: unchanged | Copilot Studio: usage-based agent execution unchanged | Azure AI Studio: promotional credits ending July 31 | Windows Copilot: free for consumer; enterprise features require M365 licensing.
Best for: Licensing admins planning August usage and credit allocation — note Azure AI Studio promo credit expiry and the August 5 Cowork limit rollback.